MAS.632 Conversational Computer Systems - MIT OpenCourseWare

120 VOICE COMMUNICATION WITH COMPUTERS
for both house number and street name and informs the user to use the "#"key
after entering the house number. Of course, explicit termination can fail if the
user neglects to use the termination key. If the user pauses, waiting for a
response, the program indicates that it thinks the caller is still entering an
address and notes that a terminator key is required. If the user forgets the terminator
and simply continues, next entering the street name, the system may
detect this. When the number of digits entered exceeds the highest known house
number in the database, the system indicates that this is an unknown address,
reminds the caller to terminate address entry with a pound sign, and starts over.
Entering a street name is more challenging due to multiple instances of the
same street name in the database. The caller is prompted to spell the name of the
street using one telephone key per character. Again explicit termination is
requested. In the case of multiple streets having identical touch tone "spellings,"
Direction Assistance resorts to a temporal menu of the possible choices. However,
much more common than spelling confusion is that the same street may be found
in multiple municipalities, or the same name can apply to multiple streets even
in one municipality (e.g., Hancock Street, Hancock Place, and Hancock Park in
Cambridge). In these circumstances, Direction Assistance again resorts to a temporal
menu to finish the selection. To help explain the selection process, the system
first indicates why it is confused, e.g., "I don't know whether you mean
Hancock Street, Place, or Park. Hit any key when you hear the one you want."
Romt. D.,4tk.
Once the origin and destination are known, the Route Finder determines a route.
The route consists of a path traversing a number of street segments (each segment
being approximately one block long). The route is optimized on the basis of
distance, quality of the streets, ease of driving (e.g., a penalty is imposed for left
turns), and ease of explanation (a zig-zag route may be shorter but may be more
complicated and difficult to follow).
The next step is the most challenging language-generation task ofthis application.
The Describer must take the series of segments from the database and turn
it into a coherent description of how to drive to the destination. A list of what to
do at each intersection is hardly useful to the driver and interferes with the goal
of speaking succinctly. Instead the Describer generates instructions comparable
to those a person gives. A passenger assumes the driver takes the "obvious" route
(e.g., stays on the main street) until told otherwise. This allows the Describer to
compress many sequential segments into a single driving act, e.g., "Stay on Main
Street until it merges with Massachusetts Avenue."
The Describer employs a taxonomy of intersection types including such classes
as continue, forced-turn, turn, fork, onto-rotary, and exit-rotary. A software
module associated with each is consulted in order at each intersection. The
appropriate module chooses to describe the driving act to perform at this intersection.
Succinctness is achieved by having the continue expert usually choose
to say nothing unless a street name changes or a significant landmark is encountered.